<s>
SHMEM	B-Operating_System
(	O
from	O
Cray	O
Research	O
's	O
“	O
shared	O
memory	O
”	O
library	O
)	O
is	O
a	O
family	O
of	O
parallel	O
programming	O
libraries	O
,	O
providing	O
one-sided	O
,	O
RDMA	B-General_Concept
,	O
parallel-processing	O
interfaces	O
for	O
low-latency	O
distributed-memory	O
supercomputers	O
.	O
</s>
<s>
The	O
SHMEM	B-Operating_System
acronym	O
was	O
subsequently	O
reverse	O
engineered	O
to	O
mean	O
"	O
Symmetric	O
Hierarchical	O
MEMory	O
”	O
.	O
</s>
<s>
Later	O
it	O
was	O
expanded	O
to	O
distributed	B-Operating_System
memory	I-Operating_System
parallel	O
computer	O
clusters	O
,	O
and	O
is	O
used	O
as	O
parallel	O
programming	O
interface	O
or	O
as	O
low-level	O
interface	O
to	O
build	O
partitioned	B-Application
global	I-Application
address	I-Application
space	I-Application
(	O
PGAS	O
)	O
systems	O
and	O
languages	O
.	O
</s>
<s>
“	O
Libsma	O
”	O
,	O
the	O
first	O
SHMEM	B-Operating_System
library	O
,	O
was	O
created	O
by	O
Richard	O
Smith	O
at	O
Cray	O
Research	O
in	O
1993	O
as	O
a	O
set	O
of	O
thin	O
interfaces	O
to	O
access	O
the	O
CRAY	O
T3D	O
's	O
inter-processor-communication	O
hardware	O
.	O
</s>
<s>
SHMEM	B-Operating_System
has	O
been	O
implemented	O
by	O
Cray	O
Research	O
,	O
SGI	O
,	O
Cray	O
Inc.	O
,	O
Quadrics	O
,	O
HP	O
,	O
GSHMEM	O
,	O
IBM	O
,	O
QLogic	O
,	O
Mellanox	O
,	O
Universities	O
of	O
Houston	O
and	O
Florida	O
;	O
there	O
is	O
also	O
open-source	O
OpenSHMEM	B-Operating_System
.	O
</s>
<s>
SHMEM	B-Operating_System
laid	O
the	O
foundations	O
for	O
low-latency	O
(	O
sub-microsecond	O
)	O
one-sided	O
communication	O
.	O
</s>
<s>
With	O
the	O
advent	O
of	O
popular	O
sub-microsecond	O
interconnects	O
,	O
SHMEM	B-Operating_System
has	O
been	O
used	O
to	O
address	O
the	O
necessity	O
of	O
hyper-efficient	O
,	O
portable	O
,	O
parallel-communication	O
methods	O
for	O
exascale	O
computing	O
.	O
</s>
<s>
Programs	O
written	O
using	O
SHMEM	B-Operating_System
can	O
be	O
started	O
on	O
several	O
computers	O
,	O
connected	O
together	O
with	O
some	O
high-performance	O
network	O
,	O
supported	O
by	O
used	O
SHMEM	B-Operating_System
library	O
.	O
</s>
<s>
Every	O
computer	O
runs	O
a	O
copy	O
of	O
a	O
program	O
(	O
SPMD	B-Operating_System
)	O
;	O
each	O
copy	O
is	O
called	O
PE	O
(	O
processing	O
element	O
)	O
.	O
</s>
<s>
PEs	O
can	O
ask	O
the	O
SHMEM	B-Operating_System
library	O
to	O
do	O
remote	O
memory-access	O
operations	O
,	O
like	O
reading	O
(	O
"	O
shmem_get	O
"	O
operation	O
)	O
or	O
writing	O
(	O
"	O
shmem_put	O
"	O
operation	O
)	O
data	O
.	O
</s>
<s>
For	O
short	O
datatypes	O
,	O
SHMEM	B-Operating_System
can	O
do	O
atomic	O
operations	O
(	O
CAS	B-Operating_System
,	O
fetch	O
and	O
add	O
,	O
atomic	O
increment	O
,	O
etc	O
.	O
)	O
</s>
<s>
SHMEM	B-Operating_System
has	O
several	O
collective	O
operations	O
,	O
which	O
should	O
be	O
started	O
by	O
all	O
PEs	O
,	O
like	O
reductions	O
,	O
broadcast	O
,	O
collect	O
.	O
</s>
<s>
OpenSHMEM	B-Operating_System
is	O
a	O
standard	O
effort	O
by	O
SGI	O
and	O
Open	O
Source	O
Software	O
Solutions	O
,	O
Inc	O
.	O
</s>
<s>
In	O
first	O
years	O
SHMEM	B-Operating_System
was	O
accessible	O
only	O
on	O
some	O
Cray	O
Research	O
machines	O
(	O
later	O
additionally	O
on	O
SGI	O
)	O
equipped	O
with	O
special	O
networks	O
,	O
limiting	O
library	O
widespread	O
and	O
being	O
vendor	O
lock-in	O
(	O
for	O
example	O
,	O
Cray	O
Research	O
recommended	O
to	O
partially	O
rewrite	O
MPI	O
programs	O
to	O
combine	O
both	O
MPI	O
and	O
shmem	B-Operating_System
calls	O
,	O
which	O
make	O
the	O
program	O
non-portable	O
to	O
other	O
clear-MPI	O
environment	O
)	O
.	O
</s>
<s>
SHMEM	B-Operating_System
was	O
not	O
defined	O
as	O
standard	O
,	O
so	O
there	O
were	O
created	O
several	O
incompatible	O
variants	O
of	O
SHMEM	B-Operating_System
libraries	O
by	O
other	O
vendors	O
.	O
</s>
<s>
Some	O
SHMEM	B-Operating_System
routines	O
were	O
designed	O
according	O
to	O
CRAY	O
T3D	O
architecture	O
limitations	O
,	O
for	O
example	O
reductions	O
and	O
broadcasts	O
could	O
be	O
started	O
only	O
on	O
subsets	O
of	O
PEs	O
with	O
size	O
being	O
power	O
of	O
two	O
.	O
</s>
<s>
Variants	O
of	O
SHMEM	B-Operating_System
libraries	O
can	O
run	O
on	O
top	O
of	O
any	O
MPI	O
library	O
,	O
even	O
when	O
a	O
cluster	O
has	O
only	O
non-rdma	O
optimized	O
Ethernet	O
,	O
however	O
the	O
performance	O
will	O
be	O
typically	O
worse	O
than	O
other	O
enhanced	O
networking	O
protocols	O
.	O
</s>
<s>
SHMEM	B-Operating_System
is	O
available	O
only	O
for	O
C	O
and	O
Fortran	O
(	O
some	O
versions	O
also	O
to	O
C++	O
)	O
.	O
</s>
<s>
Many	O
disadvantages	O
of	O
SHMEM	B-Operating_System
have	O
been	O
overcome	O
with	O
the	O
use	O
of	O
OpenSHMEM	B-Operating_System
on	O
popular	O
sub-microsecond	O
interconnects	O
,	O
driven	O
by	O
exascale	O
development	O
.	O
</s>
